Florian Juengermann × Max Agency (LangChain) — Listen agents, 2026¶
Source: YouTube · ~47min · LangChain's Max Agency podcast with harrison-chase Guest: florian-juengermann, co-founder/CTO of listen-labs Raw: transcript
Threads¶
Architecture¶
- virtual-table-architecture — rows-and-columns abstraction, not a file system; Postgres-backed
- map-reduce-classification — GPT-mini / Haiku fan-out over thousands of interviews
- powerpoint-subagent — rebuilt on claude-code SDK inside E2B
- subagent-architecture — Listen's specific flavor (extends existing page)
Context & skills¶
- contextual-prompt-engineering — their chosen alternative to skills
- skill-distillation — used selectively inside the PowerPoint subagent only
Quality & evals¶
- feedback-subagent — reviewer that doubles as eval system
- llm-as-a-judge — broader frame; Listen's reviewer is one instance
- eval-lifecycle-pre-to-production — Konstanty's lifecycle; Listen runs judge async post-chat, not live
UX¶
- composer-agent-ux — typed-edit-log co-authoring for the discussion guide
- trigger-long-running-agent — live vs async mode selection
- live-report-numbers — keeping quantitative claims accurate as new responses arrive
People & orgs¶
- florian-juengermann — CTO, Listen
- listen-labs — company
- harrison-chase — interviewer (LangChain CEO)
- langchain — show host
Cross-links into prior ingests¶
- Harness flavor: table-first, not file-first — contrasts with Lopopolo's repo-centric harness-engineering and Hashimoto's file-tree agent loop
- Subagent pattern: echoes faye-zhang-subagents-posttraining-2026 (scoped workers with own context) and harrison-sam-general-purpose-agents-2026 (Harrison's Deep Agents thesis)
- Contextual engineering vs skills: connects to dexter-horthy-qrspi-2026 and peter-steinberger-state-of-the-claw-2026 — three different bets on the same axis
- Judge-as-runtime: mahmoud-mabrouk-judge-the-judge-gepa-2026 on validating the judge itself; Listen collapses runtime review + eval into one artifact
- Eval pragmatism: maggie-konstanty-evals-2026 on eval-first-contact survival — Listen matches the "engineer reads traces, no PM layer" posture
One non-obvious cut¶
Florian's hiring pivot is the line that dates this episode: a year ago he'd hire smart non-AI engineers; by 2026 he treats "never worked with AI" as disqualifying. Lines up with software-3-0 and mihail-eric-modern-software-engineer-2026 — the baseline for agent engineers has moved.